Identification of tea plant cultivars based on canopy images using deep learning methods

• A canopy image-based dataset containing 18 cultivars for tea plant cultivar identification was constructed. • 20 tea plant cultivar identification models were developed based on deep learning methods. • The identification performance of the model for tea plant cultivar was improved by optimal training process. • The accurate and non-destructive identification of tea plant cultivars in the field conditions was achieved. Accurate and rapid identification of tea plant cultivars in field conditions contributes significantly to improve the refined and intelligent production in tea plantations. However, the high phenotypic similarity among different tea plant cultivars, coupled with the morphological variation of the same cultivar under different nutritional statuses, growth periods, and environmental conditions, poses a substantial challenge for cultivar identification. Deep learning can automatically extract deep features from images, capturing critical information about various classes. Consequently, this study employed deep learning methods to identify tea plant cultivars using canopy images and investigated the effects of different training strategies on the model training outcomes. For this purpose, canopy images of tea plant in natural environment were used as the research object in this study, and a dataset for the identification of tea plant cultivars containing 8000 canopy images of 16 cultivars was constructed. 20 tea plant cultivars identification models were developed based on convolutional neural networks and lightweight convolutional neural networks, among which DenseNet201 exhibited the highest identification performance with an accuracy of 96.38 %. In addition, the training outcomes were significantly influenced by selected optimizers, training parameters, and image augmentation methods during the model training process. The optimal training strategy involved using the AdamW optimizer to dynamically adjust the model parameters, with the learning rate, batch size, and epoch as 0.0001, 16, and 50, respectively. Additionally, the online image augmentation process with random rotation demonstrated the optimal identifying performance in tea plant cultivars, with an accuracy of 97.81 %. This study realized the accurate identification of tea plant cultivars through canopy images and verified the feasibility of using deep learning methods to identify tea plant cultivars in field conditions. This study not only provides technical support for tea producers, but also serves as a valuable reference for cultivar identification of other crops.